2020-05-04 20:51:39 +00:00
|
|
|
import math
|
2020-04-16 02:16:40 +00:00
|
|
|
from abc import ABC
|
|
|
|
from collections import OrderedDict
|
|
|
|
|
2020-05-04 20:51:39 +00:00
|
|
|
import torch
|
|
|
|
|
2020-08-06 14:58:51 +00:00
|
|
|
from pytorch_lightning import TrainResult, EvalResult
|
|
|
|
|
2020-04-16 02:16:40 +00:00
|
|
|
|
|
|
|
class TrainingStepVariations(ABC):
|
|
|
|
"""
|
|
|
|
Houses all variations of training steps
|
|
|
|
"""
|
2020-05-04 20:51:39 +00:00
|
|
|
test_step_inf_loss = float('inf')
|
|
|
|
|
2020-04-16 02:16:40 +00:00
|
|
|
def training_step(self, batch, batch_idx, optimizer_idx=None):
|
|
|
|
"""Lightning calls this inside the training loop"""
|
|
|
|
# forward pass
|
|
|
|
x, y = batch
|
|
|
|
x = x.view(x.size(0), -1)
|
|
|
|
|
|
|
|
y_hat = self(x)
|
|
|
|
|
|
|
|
# calculate loss
|
|
|
|
loss_val = self.loss(y, y_hat)
|
|
|
|
|
|
|
|
# alternate possible outputs to test
|
2020-05-31 12:29:51 +00:00
|
|
|
output = OrderedDict({
|
|
|
|
'loss': loss_val,
|
|
|
|
'progress_bar': {'some_val': loss_val * loss_val},
|
|
|
|
'log': {'train_some_val': loss_val * loss_val},
|
|
|
|
})
|
|
|
|
return output
|
2020-05-04 20:51:39 +00:00
|
|
|
|
|
|
|
def training_step__inf_loss(self, batch, batch_idx, optimizer_idx=None):
|
|
|
|
output = self.training_step(batch, batch_idx, optimizer_idx)
|
|
|
|
if batch_idx == self.test_step_inf_loss:
|
|
|
|
if isinstance(output, dict):
|
|
|
|
output['loss'] *= torch.tensor(math.inf) # make loss infinite
|
|
|
|
else:
|
|
|
|
output /= 0
|
|
|
|
return output
|
2020-07-20 23:00:20 +00:00
|
|
|
|
|
|
|
def training_step_full_loop_result_obj_dp(self, batch, batch_idx, optimizer_idx=None):
|
|
|
|
"""
|
|
|
|
Full loop flow train step (result obj + dp)
|
|
|
|
"""
|
|
|
|
x, y = batch
|
|
|
|
x = x.view(x.size(0), -1)
|
|
|
|
y_hat = self(x.to(self.device))
|
|
|
|
loss_val = y_hat.sum()
|
|
|
|
result = TrainResult(minimize=loss_val)
|
|
|
|
result.log('train_step_metric', loss_val + 1)
|
|
|
|
self.training_step_called = True
|
|
|
|
return result
|
|
|
|
|
|
|
|
def training_step_end_full_loop_result_obj_dp(self, result):
|
|
|
|
"""
|
|
|
|
Full loop flow train step (result obj + dp)
|
|
|
|
"""
|
|
|
|
result.minimize = result.minimize.mean()
|
|
|
|
result.checkpoint_on = result.checkpoint_on.mean()
|
|
|
|
result.train_step_metric = result.train_step_metric.mean()
|
|
|
|
result.log('train_step_end_metric', 1)
|
|
|
|
self.training_step_end_called = True
|
|
|
|
return result
|
|
|
|
|
|
|
|
def training_epoch_end_full_loop_result_obj_dp(self, result):
|
|
|
|
"""
|
|
|
|
Full loop flow train step (result obj + dp)
|
|
|
|
"""
|
|
|
|
result.log('train_epoch_end_metric', 1, on_epoch=True)
|
|
|
|
self.training_epoch_end_called = True
|
|
|
|
return result
|
2020-07-22 17:53:10 +00:00
|
|
|
|
|
|
|
def eval_step_full_loop_result_obj_dp(self, batch, batch_idx, optimizer_idx=None):
|
|
|
|
"""
|
|
|
|
Full loop flow train step (result obj + dp)
|
|
|
|
"""
|
|
|
|
x, y = batch
|
|
|
|
x = x.view(x.size(0), -1)
|
|
|
|
y_hat = self(x.to(self.device))
|
|
|
|
loss_val = y_hat.sum()
|
|
|
|
result = EvalResult(checkpoint_on=loss_val, early_stop_on=loss_val)
|
|
|
|
|
|
|
|
eval_name = 'validation' if not self.trainer.testing else 'test'
|
|
|
|
result.log(f'{eval_name}_step_metric', loss_val + 1, on_step=True)
|
|
|
|
|
|
|
|
setattr(self, f'{eval_name}_step_called', True)
|
|
|
|
return result
|
|
|
|
|
|
|
|
def eval_step_end_full_loop_result_obj_dp(self, result):
|
|
|
|
"""
|
|
|
|
Full loop flow train step (result obj + dp)
|
|
|
|
"""
|
|
|
|
eval_name = 'validation' if not self.trainer.testing else 'test'
|
|
|
|
reduced = getattr(result, f'step_{eval_name}_step_metric').mean()
|
|
|
|
setattr(result, f'step_{eval_name}_step_metric', reduced)
|
|
|
|
|
|
|
|
reduced = getattr(result, f'epoch_{eval_name}_step_metric').mean()
|
|
|
|
setattr(result, f'epoch_{eval_name}_step_metric', reduced)
|
|
|
|
|
|
|
|
result.checkpoint_on = result.checkpoint_on.mean()
|
|
|
|
result.early_stop_on = result.early_stop_on.mean()
|
|
|
|
result.log(f'{eval_name}_step_end_metric', torch.tensor(1).type_as(result.checkpoint_on))
|
|
|
|
setattr(self, f'{eval_name}_step_end_called', True)
|
|
|
|
return result
|
|
|
|
|
|
|
|
def eval_epoch_end_full_loop_result_obj_dp(self, result):
|
|
|
|
"""
|
|
|
|
Full loop flow train step (result obj + dp)
|
|
|
|
"""
|
|
|
|
eval_name = 'validation' if not self.trainer.testing else 'test'
|
|
|
|
result.log(f'{eval_name}_epoch_end_metric', torch.tensor(1).type_as(result.checkpoint_on), on_epoch=True)
|
|
|
|
result.checkpoint_on = result.checkpoint_on.mean()
|
|
|
|
result.early_stop_on = result.early_stop_on.mean()
|
|
|
|
setattr(self, f'{eval_name}_epoch_end_called', True)
|
|
|
|
|
|
|
|
# reduce the parametrized values
|
|
|
|
reduced = getattr(result, f'step_{eval_name}_step_metric').mean()
|
|
|
|
setattr(result, f'step_{eval_name}_step_metric', reduced)
|
|
|
|
|
|
|
|
reduced = getattr(result, f'epoch_{eval_name}_step_metric').mean()
|
|
|
|
setattr(result, f'epoch_{eval_name}_step_metric', reduced)
|
|
|
|
|
|
|
|
reduced = getattr(result, f'{eval_name}_step_end_metric').mean()
|
|
|
|
setattr(result, f'{eval_name}_step_end_metric', reduced)
|
|
|
|
|
|
|
|
return result
|